Viliana Devbunova

Yandex, Research ML Engineer

Viliana Devbunova is a Research ML Engineer on the speech and language team at Yandex, where she has spent four years building production speech systems — neural text-to-speech pipelines, including vocoders and codec-based synthesis with self-supervised speech tokenization, as well as streaming speech recognition and real-time translation. Her work spans the full path from research prototypes to systems serving millions of users.

In parallel, she pursues research on LLM controllability and mechanistic interpretability, focusing on detecting behaviors such as sycophancy and evaluation awareness via linear probes on model internals. Her work on the limitations of probe-based evidence for evaluation awareness was presented at an ICLR 2026 workshop. She is a graduate of the ARENA program and Neel Nanda’s AI Safety Training Program.

Viliana holds BSc and MSc degrees with honors from the Faculty of Computational Mathematics and Cybernetics at Lomonosov Moscow State University. At PSIML she will give a lecture and a hands-on seminar on text-to-speech.